基於智慧型手機上的時間空間資訊進行使用者行為推測

Abstract

隨著生活中收集地理位置相關資訊的移動設備的普及以及越來 越多的基於位置的服務收集所得到的豐盛的位置數據，使得不使用連 續收集的感應器數據來建立個人化的使用者行為探測成為可能。可以 發現用戶在使用智慧型手機進行拍照，打卡，接入WiFi 等，這些行 為時會產生足以捕捉用戶行為的時間空間數據。因此，在本文中，我 們只需要通過從手機上收集到的個人空間時間的習慣以及所處地方 的大眾行為訊息就可以打造一個個性化的模型來推斷用戶的行為。通 過在智能手機上分析這些時間空間的數據，有三種類型的特徵可以用 於使用者行為的推測：1）地理位置特徵：捕捉使用者日常行為的地 理活動; 2）時間特徵：它反應了用戶行為的時間規律; 3）位置的語義 特徵：從基於位置的社交網絡服務獲得大眾對每個地方的功能以及適 宜從事的活動的意見。我們從收集到的時間空間數據中對這三種特徵 進行勘探、融合，進而提出了一個基於時間空間數據建立使用者行為 推測模型的框架STAIM。它可以根據用戶的行為，動態地確定權重。 我們進行了大量基於真實數據的實驗并與現有的類似解法相比，結果 證明STAIM 能夠僅根據來自智能手機的時間空間數據推斷用戶的行 為。同時，也通過實驗對各參數的敏感性進行了分析，STAIM 可以 根據用戶的歷史行為建立一個個性化的行為推測模型。

With the plenteous location data collected from commodity mobile devices and widespread location-based services, it is possible to infer user activities without continuously monitoring context information from sensors in mobile phones. Note that users take photos, perform check-in behaviors and access WiFi networks in their smart phones. These behaviors generate spatial-temporal data that capture user activities. Thus, in this paper, we only take advantage of the individual spatial-temporal habit and the public opinions of visited places, and build a personalized model to infer user activities. By analyzing these spatial-temporal data in smart phones, there are three types of features used for activity inferences: 1) the geographical feature which captures the daily movement behaviors accompanied with activities; 2) the temporal feature which records the individual routine schedule of activities; 3) the semantic feature which is obtained from location-based social network and stands for the view of the place taking a function once performed by volunteers on behalf of majority. In light of three types of features hidden in spatial-temporal data, we fuse the concerned 3 features together and develop an activity inference framework named Spatial-Temporal Activity Inference Model (STAIM), which could dynamically determine the weights according to users' behavior. By comparing with the state-of-art and other existing works over the real data set, the empirical results elucidate that STAIM is able to infer user activities without collecting more sensing readings from smart phones. Sensitive analysis on parameters is conducted and STAIM could achieve personalized models based on users' behavior.